From Idea to Working App in Days, Not Months

I built an Android keyboard. Open-source, privacy-first, with built-in voice input via Mistral's Voxtral ASR. It's called Ownkey and you can download it now.

That might sound like a side project that took months. It wasn't.

It started with a model that felt different

I test a lot of AI models. Most are impressive in a demo, but the moment you put them in a real workflow, the magic fades. Latency just a bit too high. Output just slightly off. You keep correcting manually.

Voxtral Mini Realtime was different. Not spectacularly different. Just usable. Fast enough that you don't have to wait for it. Accurate enough that you don't constantly need to fix the output. That difference between "almost" and "good enough" matters more than you'd think.

At the same time, I was using Wispr Flow and saw how natural voice-to-text can be when the friction is low enough. That combination, a fast ASR model and the experience that speech as input just works, was the direct trigger. This needs to be in a keyboard. Not as a separate app, but right where you're already typing.

Why a keyboard

Most AI tools live in their own app. You have to switch to them, provide context, copy the output, paste it back. That's friction. A keyboard is everywhere. In WhatsApp, in your email, in Slack. No context-switch needed.

Ownkey solves one specific problem: voice-to-text you can trust, directly in your keyboard, with your own API key. No platform storing your data. No subscription. You run your own configuration.

Here's what's in it:

• Mistral Voxtral ASR as voice input, directly in the keyboard's dictation flow (no external IME switch)
• Bring Your Own Key: you enter your Mistral API key, stored via Android Keystore-backed encrypted preferences
• Autocorrect tuned for real typing behavior, with EN/NL frequency dictionaries
• Dark theme with high contrast, Samsung-inspired quick actions, adjustable corner radius

It's a fork of FlorisBoard. Not a from-scratch rewrite, deliberately so. Take a solid foundation and build specifically on top of it. That's how you move fast without sacrificing quality.

The real accelerator: an agent that builds while you do something else

Ownkey was built in a single day. Not by me behind an IDE. By OpenClaw, directed via Telegram, running on GPT 5.3 Codex.

I set the direction. The problem, the architecture decisions, the priorities. OpenClaw did the building: writing code, iterating, testing, committing. Via Telegram, while I was busy with other things.

That's a fundamentally different model than "AI as assistant." This isn't autocomplete on steroids. This is an agent that independently builds an app based on your specs, and keeps you updated on progress via chat. You steer when needed. The agent delivers.

One day. From idea to working Android app. While I was doing something else.

What I'm learning from this

The interesting shift with AI isn't in generating more ideas. Ideas are cheap. The shift is in the distance between idea and working product.

A year ago, Ownkey would have been a side project sitting somewhere on a backlog. Now it's an app built in a single day, with an open-source repo, documentation, and a roadmap.

That's not because of one magical tool. It's because of a combination: a model that's fast enough (Voxtral), the inspiration from a product that shows how low friction can be (Wispr Flow), and an agent that doesn't just think along but actually builds. OpenClaw with GPT 5.3 Codex, directed via Telegram.

The win isn't in the tool itself. The win is in how little friction remains between thinking and building.

---

Ownkey is available for download on Android now. The repo is open: github.com/MajesteitBart/ownkey-keyboard